llvm dev - Jun 2017 - Possible llc bug: "Virtual register defs don't dominate all uses."

Hello.
     I am experiencing some errors in the back end (which I presume are actually
a bug of
lib/CodeGen/MachineVerifier.cpp) such as:
       *** Bad machine code: Virtual register defs don't dominate all uses.
***
       - function:    MatMul_RowMajor_Transposed
       - v. register: %vreg93

       *** Bad machine code: Virtual register defs don't dominate all uses.
***
       - function:    MatMul_RowMajor_Transposed
       - v. register: %vreg88
       LLVM ERROR: Found 2 machine code errors.

     These errors are given because of the following machine basic block:
       BB#5: derived from LLVM BB %middle.block
         Predecessors according to CFG: BB#4
           %vreg88<def,dead> = LDSH_H
           ; IMPORTANT: here an instruction bundle starts
           %vreg92<def,dead> = EQ_H [...]
             * NOP
             * WHEREEQ
             * %vreg93<def> = VLOAD_H 0
             * %vreg87<def> = ORV_H %vreg86, %vreg93
             * %vreg89<def> = ORV_H %vreg88, %vreg93
           ; IMPORTANT: here the instruction bundle ends
           ENDWHERE
           RED_H %vreg87
           RED_H %vreg89

     Basically the problem is related to the fact I have created an instruction
bundle in
which I define and use virtual register %vreg93 and also I use %vreg88 defined
above it
(note that if I do NOT use inside the bundle a virtual register I do not get
this error).

     Did anybody else encounter something like this? It looks to me this is
actually a bug
of lib/CodeGen/MachineVerifier.cpp. Can you tell me how can I address this
problem?

   Thank you,
     Alex

On 06/05/2017 08:17 AM, Alex Susu via llvm-dev wrote:
>   Hello.
>     I am experiencing some errors in the back end (which I presume are
actually a bug of lib/CodeGen/MachineVerifier.cpp) such as:
>       *** Bad machine code: Virtual register defs don't dominate all
uses. ***
>       - function:    MatMul_RowMajor_Transposed
>       - v. register: %vreg93
> 
>       *** Bad machine code: Virtual register defs don't dominate all
uses. ***
>       - function:    MatMul_RowMajor_Transposed
>       - v. register: %vreg88
>       LLVM ERROR: Found 2 machine code errors.
> 
>     These errors are given because of the following machine basic block:
>       BB#5: derived from LLVM BB %middle.block
>         Predecessors according to CFG: BB#4
>           %vreg88<def,dead> = LDSH_H
>           ; IMPORTANT: here an instruction bundle starts
>           %vreg92<def,dead> = EQ_H [...]
>             * NOP
>             * WHEREEQ
>             * %vreg93<def> = VLOAD_H 0
>             * %vreg87<def> = ORV_H %vreg86, %vreg93
>             * %vreg89<def> = ORV_H %vreg88, %vreg93
You should be setting the InternalRead flag for the vreg93 uses.
I think that will fix the first error.

-Tom
>           ; IMPORTANT: here the instruction bundle ends
>           ENDWHERE
>           RED_H %vreg87
>           RED_H %vreg89
> 
>     Basically the problem is related to the fact I have created an
instruction bundle in which I define and use virtual register %vreg93 and also I
use %vreg88 defined above it (note that if I do NOT use inside the bundle a
virtual register I do not get this error).
> 
>     Did anybody else encounter something like this? It looks to me this is
actually a bug of lib/CodeGen/MachineVerifier.cpp. Can you tell me how can I
address this problem?
> 
>   Thank you,
>     Alex
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev